Exclusive For those of us on the outside of Sun Microsystems, the future of the Sparc processor and its related server platforms has been the subject of much speculation and debate. But for Sun's largest customers, the mystery has been over since sometime in June.
That's when Sun's execs made the rounds to the biggest Sparc shops in the world and gave them a preview of the Sparc roadmap over the next three years. Under non-disclosure, of course. Like that was going to hold, considering all of the curiosity in the wake of Oracle's pending $7.4bn (net around $5.2bn) acquisition of Sun.
El Reg has obtained the one-page roadmap that Sun's customers - and presumably also Fujitsu's customers - have been shown about the future Sparc processor lineup. As we originally reported in mid-June, the 16-core "Rock" UltraSparc-RK processor for Sun's once-and-never "Supernova" line of servers is not on the roadmap.
Sun has never confirmed rumors that Rock was dead, which we heard second-hand from people who heard it from Sun sources who know. Then in early August, the tweaks for the Rock chips were removed from the OpenSolaris development version of the Solaris Unix variant, which referred to Rock as the UltraSparc-AT10. Call it what you will, but most Sun watchers think Rock is dead, even if Sun isn't (officially) talking.
The roadmap above is by no means the ultimate roadmap that Oracle will use to pilot the Sun server business once it has acquired the company, which it almost surely will because no one else wants Sun, at least not in its entirety. And despite all of its bravado about taking on IBM in the server business in recent weeks (see here and here), it would be hard to find anyone in the IT business who wasn't skeptical about Oracle's long-term commitment to being in the relatively low margin server biz, especially since it not only apparently tried to sell off Sun's server biz to Hewlett-Packard before launching its acquisition, but also tried to sell it off since the proposed takeover was announced in April. Oracle has not confirmed any of this, of course.
Let's get back to the roadmap (below), starting on top because it is easier. We can expect a goosed Sparc64-VII+ chip any day now, which will run at 2.88 GHz and which will be a four-core, eight-threaded chip like its "Jupiter" predecessor. This Jupiter+ chip is implemented in the same 65 nanometer process as the Jupiter chip was, and it is made by Fujitsu, a company that is in the process of outsourcing its chip manufacturing to Taiwan Semiconductor Manufacturing Corp. A long way off in late 2010 or early 2011, the Sparc Enterprise server lineup gets a speed boost to 3 GHz with the Jupiter-E chips.
After that, in 2012, Sun has made no commitment to the kicker line of Fujitsu "Advanced Product Line 2" servers coming from Fujitsu. These APL2 machines are presumably to be based on the "Venus" eight-core Sparc64-VIII processor, which has a Sparc64-VIIIfx variant aimed at supercomputers. That Sparc64-VIIIfx chip will be used in a 10 petaflops massively parallel machine being built by Fujitsu and paid for by the Japanese government under the 1.2bn Project Keisoku effort.
(This project originally called for the three key server makers in Japan - NEC, Hitachi, and Fujitsu - to cooperate in building a 10-petaflop machine made up of scalar processors designed by Fujitsu, vector processors designed by NEC and Hitachi, and a system design created by all three. Earlier this year, NEC and Hitachi pulled out and in July, Fujitsu took over the whole contract).
There's no 16-core, 2.1 GHz UltraSparc-RK chip here. No box that is going to offer 30 times more oomph on "data throughput" workloads compared to systems using 1.2 GHz UltraSparc-III processors, as Sun was promising back in 2005 through 2007. And with Oracle being a big Sun shop itself, it would be funny to learn if Oracle's own IT shop is a bit miffed at Sun and Fujitsu at this point, after many product delays and now, the death of Rock.
Please help me. I'm falling...
That leaves us with the Niagara family of chips, known as the Sparc T series outside of Sun and used in entry and midrange servers. El Reg has already told you about the faster Niagara-2+ (T2) and Victoria Falls+ (T2+) processors that received clock speed boosts in July.
A few weeks ago, at the Hot Chips 21 conference, Sun talked very generally about the 16-core "Rainbow Falls" Sparc T series processor, which is known variously as the Niagara-3 and the KT chip if you are having trouble keeping track. It looks like this future Sparc T chip has been changed a few times. Back in June 2008, El Reg learned that this future Niagara family chip would sport 16 cores and 16 threads per core, for a total of 256 threads per socket.
That's as many threads as Sun can cram into a four-socket T5440 server today using its Sparc T2+ chips. The servers based on this chip were supposed to scale to as high as eight sockets in a single system image, for a total of 2,048 threads. When Sun talked about Rainbow Falls at Hot Chips at the end of August, it said the chips would have 16 cores but didn't specify the thread count, like you do on sheets. The cryptographic accelerators that Sun talked about in a separate presentation said that they had eight "strands," and I didn't see that this also meant that the next generation of Sparc T chips would only have eight threads per core, just like the T2 and T2+ already do.
What we learn from the Sparc roadmap above is that not only has Sun cut back on the threads with Rainbow Falls, it has also cut back on the socket count, keeping it at the same four sockets used by the T5440 server. And instead of hitting something close to 2 GHz as it should be able to do as it shifts from a 65 nanometer to a 45 nanometer process in the middle of 2010, Sun is only telling customers that it can boost clock speeds to 1.67 GHz with Rainbow Falls.
This, to use a technical term in the computer business, sucks.
With the "Yosemite Falls" kicker to the T3 chip, presumably to be sold as the T4, Sun is planning to implement a new core for the Niagara family, and it will backstep to eight cores with eight threads per core. If you are thinking this is necessary to boost clock speeds, you are right, since this chip will use the same 40 nanometer process as Rainbow Falls T3s and will run at 2.5 GHz. However, the SMP scalability of the chip will still be limited to a paltry four sockets.
In the middle of 2012, something interesting is on the roadmap, a chip code-named "Yellowstone Falls", which has only four cores and eight threads per core that runs at 3 GHz. This chip will be implemented in a 28 nanometer process and will be used in servers that span from 4 to 192 sockets. (Yes, I said 192, as you can see from the pretty picture). At the end of 2012, there is a related chip called "Cascade Falls" that runs at the same 3 GHz and has 16 cores and eight threads per core like the future Rainbow Falls T3 chips, but it can be used in machines with anywhere from 1 to 8 processor sockets.
All of this is subject to change, and some of it most certainly will once Oracle takes control of Sun. ®